STRATUM is a Horizon Europe multi-partner project developing a clinically validated, real-time 3D decision support tool for brain tumour surgery. The system integrates Hyperspectral Imaging (HSI), AI-based multimodal data fusion, and heterogeneous High-Performance Computing (HPC) architectures combining Graphics Processing Units (GPUs), Field-Programmable Gate Arrays (FPGAs), and Processing-In-Memory (PIM) technologies. A touchless augmented reality interface facilitates safe and intuitive intraoperative interaction. The distinguishing characteristic of STRATUM is its end-to-end co-designed approach, which integrates advanced computing, state-of-the-art imaging and clinical expertise into a unified Point-of-Care (PoC) platform. Utilising a structured co-creation methodology involving surgeons, engineers, and social scientists, the project ensures usability, safety and regulatory compliance from its early design stages to its clinical validation. The usability of STRATUM will be tested in three hospitals located in different European regions with diverse conditions and regulations. This will allow to collect advice and remarks from surgical staff in a continuous co-creation and co-tuning protocol. Beyond its clinical objectives, STRATUM contributes to the advancement of heterogeneous computing for real-time diagnostics, AI acceleration in critical medical environments and energy-efficient system integration. Furthermore, it delivers open datasets, validated AI pipelines, and performance benchmarks with a view to fostering future research and industrial innovation in digital surgery. The STRATUM project establishes a replicable model for intelligent, human-centred computing integrating microelectronics, AI and medicine.The paper presents an overview of the project in terms of aims, concepts and technologies and the description of the state of the work when approaching the end of the second of the five years planned. Specifically, the outcomes of the steps related to the collaboration with surgeons and medical staff (co-creation process) and the intelligent Graphical User Interface (GUI) development will be described. The latter allows for contactless interaction of the surgeon with several functions that have already been developed in the system.

Real-time phonocardiogram analysis on embedded devices is a key enabler for scalable and accessible cardiovascular diagnostics, particularly considering portable systems designed for low-income countries. This work introduces a combination of compressive sensing and deep learning to build portable, efficient and effective diagnostic tools for widespread cardiac screenings. The proposed classification framework is tailored for deployment on ultra-low power STM32 microcontrollers equipped with the novel Neural-ART accelerator. Experimental evaluation on the CirCor Digiscope Phonocardiogram dataset demonstrates that even with a compression ratio exceeding 100×, a classification model can achieve up to 97.3% F1-score. A similar level of performance was obtained on the PhysioNet 2016 dataset, which was used to assess the robustness and generalization capability of the developed architectures. Compared to the state-of-the-art, our final edge solution achieves 94.3% accuracy, an inference time of 18.7 ms and an energy requirement of just 1.51 mJ per input window of 4,096 samples, confirming its suitability for real-time, energy-constrained medical applications.

Integrated digital diagnostics can support complex surgical procedures in many anatomical sites, brain tumour surgery being the most complex. STRATUM is a 5-year Horizon Europe funded project with the goal of developing an innovative 3D decision support tool for brain tumour surgeries, based on real-time multimodal data processing using artificial intelligence algorithms. The proposed tool is envisioned as an energy-efficient Point-of-Care computing system to be integrated within neurosurgical workflows to aid surgeons to make informed, efficient, and accurate decisions during surgical procedures. The expected long-term impact of STRATUM is to reduce the duration of surgical procedures, thus decreasing patients' risks, but also optimising the resources of European health care systems.

Hyperspectral imaging (HSI) captures detailed spectral information across numerous wavelengths, providing superior object characterization to conventional RGB imaging. Despite these advantages, training deep learning models on HSI data is challenging due to the limited availability of extensive datasets, unlike the more familiar RGB images. To address this issue, we propose an encoder model that transforms hyperspectral images into enriched RGB images. These new enriched images represent a graphical depiction of HSI and become a new dataset to use as input for well-known models pre-trained on RGB images. In this work, we introduce HS2RGB, an encoder model based on the Vision Transformer (ViT) architecture, which condenses hyperspectral data into a three-element vector interpreted as RGB channels. The results demonstrate the effectiveness of the new images generated by the encoder, showing better visual differentiation of features compared to traditional RGB images. Morover, results highlighted greater consistency in latent vectors of the same type of tissue across different samples compared to images generated with feature selection and transformation techniques like PCA and t-SNE. Finally, we tested the enriched RGB images using Meta's SAM model for instance segmentation, revealing that our model's images provided more precise identification of regions of interest, such as tumours in medical images.

Cardiovascular Diseases (CVDs) stand as the leading cause of mortality worldwide. Detecting subtle heart sounds alterations in the early stages of CVDs can be crucial for an initial effective treatment. Accordingly, the analysis of Phonocardiograms (PCGs) through segmentation could be helpful for CVDs screening. A well-established algorithm for this task is based on a Convolutional Neural Network (CNN) with an encoding-decoding topology. Prior to the CNN processing., a computationally intensive input pre-processing., based on envelopes extraction, is needed. Thus., achieving real-time performance can be challenging. The main goal of this study is the hardware design., implementation, and evaluation of four PCG pre-processing circuits to be employed together in the design of a low-power point-of-care device for real-time Artificial Intelligence (AI)-based PCG segmentation. Results have shown that the approximations introduced by the fixed-point format and this innovative architecture have a negligible impact on the AI segmentation quality. Finally, the pre-processing chain is real-time compliant., achieving a maximum latency of 257 ms for an available processing window of 1.28 s., while dissipating only 61 mW of power.

Abstract Hyperspectral imaging is applied in the medical field for automated diagnosis of diseases, especially cancer. Among the various classification algorithms, the most suitable ones are machine and deep learning techniques. In particular, Vision Transformers represent an innovative deep architecture to classify skin cancers through hyperspectral images. However, such methodologies are computationally intensive, requiring parallel solutions to ensure fast classification. In this paper, a parallel Vision Transformer is evaluated exploiting technologies in the context of Edge and Cloud Computing, envisioning portable instruments’ development through the analysis of significant parameters, like processing times, power consumption and communication latency, where applicable. A low-power GPU, different models of desktop GPUs and a GPU for scientific computing were used. Cloud solutions show lower processing times, while Edge boards based on GPU feature the lowest energy consumption, thus resulting as the optimal choice regarding portable instrumentation with no compelling time constraints.

Cardiovascular diseases (CVDs) are currently one of the leading causes of death worldwide. Being able to detect their symptoms at early stages, even the most hidden ones, is crucial to shorten the diagnosis time and facilitate an early treatment. Currently, the use of continuous tracking systems, mainly based on wearable devices that analyze data using artificial intelligence (AI) algorithms, is being explored to automatically identify, in real time, CVDs symptoms. This could be especially relevant in lowincome countries where there is a shortage of specialized doctors. Therefore, this work focuses on analyzing the real-time execution of the state-of-the-art convolutional neural network (CNN) for heart sound segmentation (HSS) on platforms such as traditional CPU/GPU and the Fraunhofer IMS © AIRISC Core Complex (a RISC-V processor developed for AI). Results revealed that, while all implementations exploiting the CPU/GPU platform proved to be useful in real-time diagnosis from a fixed location, the AIRISC demonstrated its goodness, as a system on a chip (SoC) for a real-time wearable application, when executing a quantized version of the CNN.

Recently, several medical applications have relied on hyperspectral imaging. This technology enables both automated diagnosis and surgeon guidance. The employed algorithms adopt machine and deep learning methods to classify the images. In particular, Vision Transformers are a recent deep architecture that have been used to classify hyperspectral images of skin cancers achieving interesting results. However, deep architectures are computationally intensive and parallel architectures are mandatory to ensure a fast classification (depending on the application type even in real time). In this paper, we propose a parallel Vision Transformer architecture exploiting a low power GPU targeting the development of a portable diagnostic device. The classification time and power consumption of the low power board are compared with the performance of a desktop GPU. The results clearly highlight the suitability of the low power GPU to develop a portable diagnostic system based on hyperspectral imaging.

Among the different contactless techniques for medical diagnosis, hyperspectral imaging has gained relevance due to the high accuracy in tissues classification. Several techniques have been proposed to elaborate these images, ranging from traditional machine learning methods to deep learning algorithms. This paper evaluates three popular machine learning methods, namely Support Vector Machine (SVM), Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) considering a dataset of hyperspectral skin cancer images. The study demonstrates that the proposed algorithms are suitable for medical hyperspectral data classification, particularly when considering a small dataset.

In recent years, hyperspectral imaging has been employed in several medical applications, targeting automatic diagnosis of different diseases. These images showed good performance in identifying different types of cancers. Among the methods used for classification, machine learning and deep learning techniques emerged as the most suitable algorithms to handle these data. In this paper, we propose a novel hyperspectral image classification architecture exploiting Vision Transformers. We validated the method on a real hyperspectral dataset containing 76 skin cancer images. Obtained results clearly highlight that the Vision Transforms are a suitable architecture for this task. Measured results outperform the state-of-the-art both in terms of false negative rates and of processing times. Finally, the attention mechanism is evaluated for the first time on medical hyperspectral images.

Automatic classification of time series signals acquired by wearable or portable devices covers a central role in many critical healthcare applications, such as heart rate monitoring [1], sleep apnea study [2], gait analysis [3] and fall detection [4]. In recent years, many approaches have been adopted, including a wide range of methods ranging from threshold-based algorithms to Deep Learning techniques. The threshold-based methods have the advantage of being simple and not heavy from a computational point of view, but at the cost of low accuracy. Deep Learning approaches ensure a higher precision, but the computational complexity is increased. This is a critical issue for wearable devices because a high computational complexity strongly affects the processing time and the battery life. In this paper, we propose a hardware architecture for time series analysis using Recurrent Neural Networks (RNNs) exploiting FPGA technology. The architecture is validated with three-axial accelerometer data acquired by a wearable device used for automatic fall detection. The experimental results show that the proposed architecture outperforms state of the art solutions both in terms of processing time and power consumption.

Raman imaging is a hyperspectral approach able to provide information on the spatial distribution of a particular biochemical feature without the use of any staining or sample processing. The extraction of the relevant information from the large dataset obtained however is a laborious and complex task that still requires the development of robust chemometric approaches. In this paper, we propose a general framework for analyzing data acquired by a commercial Raman spectrometers. This framework is based both on exploiting spectral information and unsupervised clustering, in order to clearly identify the borders and the compositions of different regions of interest. Finally, we describe an efficient GPU-based parallelization, which ensures a fast image classification.

High performance computing (HPC) is becoming mandatory for the simulation of complex and realistic neuronal models. The development of such realistic models will allow to discover innovative therapies and to study brain diseases without undertaking invasive experiments that are not always possible. However, the models complexity requires adopting suitable technologies in order to provide results in short times, hopefully in real-time. To address this issue, the authors decided to exploit Graphics Processing Units (GPUs) in order to develop a realistic and morphologically detailed Purkinje cell model. This paper describes the simulation of the Purkinje cell activity adopting both single and multi-GPU strategy, together with the exploitation of different NVIDIA architectures. Results shows that the simulation times of 10000 cells is reduced from 13 days and 18 hours to about 2 hours.

Unintentional falls are the leading cause of fatal injuries and nonfatal trauma among older adults. An automated monitoring system that detects occurring falls and issues remote notifications will prove very valuable for improving the level of care that could be provided to people at higher risk. The work presented focuses on the design of embedded software for wearable devices that are connected in wireless mode to a remote monitoring system. The work focuses on the implementation of recurrent neural networks (RNNs) architectures of micro controller units (MCU) for fall detection with tri-axial accelerometers. A few general formulas for determining memory, computing power and power consumption for such architectures are presented. These formulas have been validated with an actual implementation for the SensorTile device by STMicroelectronics.

This paper describes the challenge of real-time tumor tissue identification dealt with by the HypErspectraL Imaging Cancer Detection (HELICoiD) European project. This project was funded by the Research Executive Agency, through the Future and Emerging Technologies (FET-Open) programme, under the 7th Framework Programme of the European Union. It involved four universities, three industrial partners and two hospitals. In this paper, we focused on the activity performed by the University of Las Palmas de Gran Canaria, in collaboration with the University of Pavia, concerning the parallel implementation of Support Vector Machine (SVM) classification for tumor tissue identification during surgery. Obtained results show that this classification is real-time compliant when performed using Graphic Processing Units (GPUs).

The use of High Performance Computing (HPC) technologies is gaining interest in the field of neuronal activity simulations. In fact, scientists' main goal is to understand and reproduce cells behavior in a realistic way. This will allow undertaking in silico experiments, instead of in vivo ones, to test new medicines, to study cerebral pathologies and to discover innovative therapies. To this aim, two main requirements are necessary: neurons have to be described by realistic models and their simulation hopefully have to satisfy the real-time constraint. This last property is very hard to accommodate because models used in these works are very heavy from the computational point of view. For this reason, authors decide to exploit Graphic Processing Unit (GPU) technology to simulate the cellular activity of Golgi cells, which constitute the cerebellar cortex. This paper describes an efficient Golgi cell activity simulation performed using NVIDIA GPUs. Results show that simulation times are reduced from 41 hours to about 2 hours when simulating 400'000 different cells.

This paper describes the challenge of understanding brain function through high performance computing dealt with by the Human Brain Project, the European Commission Future and Emerging Technologies Flagship involving a consortium of 112 partners spread in 24 European countries. In particular, we describe the activity, performed by one of the Italian units involved into the project, aiming at identifying very accurate models of cerebellum neurons. These models are processed through high end Graphic Processing Units (GPUs) during the tuning phase and later implemented on FPGA-based application specific processors for respecting real time requirements together with embedded implantability. Models and performance of granular neurons implementations are given in the results section.

In this letter, we present a new parallel implementation of the vertex component analysis (VCA) algorithm for spectral unmixing of remotely sensed hyperspectral data on commodity graphics processing units. We first developed a C serial version of the VCA algorithm and three parallel versions: one using NVIDIA's Compute Unified Device Architecture (CUDA), another using CUDA basic linear algebra subroutines library CUBLAS, and the last using the CUDA linear algebra library CULA. Experimental results, based on the analysis of hyperspectral images acquired by a variety of hyperspectral imaging sensors, show the effectiveness of our implementation, which satisfies the real-time constraints given by the data acquisition rate.